Visual Tracking Via Local Coordinate Coding

Visual object tracking is a basic problem in the field of computer vision. It has been a hot research topic in the field of computer vision. The observation model of the visual object is an important part of the visual object tracking problem, and a lot of methods have been put forward. Among these numerous observation models, the discriminative model (e.g. AdaBoost, SVM) is a very effective branch, this kind of method respectively uses object and background as positive and negative training samples to train a classifier as the observation model. This model not only uses the characteristics of the target itself, but also adds the background information. This is an effective model for visual object tracking since in the process of tracking due to illumination changes, deformation and occlusion factors the object usually has great changes. However, although the target itself has great changes, the target and the background will still have significant differences.In practice, the target and background feature distribution in space is usually nonlinear, that is we need to construct a nonlinear observation model to separate the target and background effectively. Therefore, although the discriminative model is an effective tracking model, a good nonlinear discriminative model is not easy to construct. In this paper, we propose an effective nonlinear classifier based on local coordinate coding method as observation model. Local coordinate coding is able to transform the global high dimensional nonlinear function into a linear combination of local linear functions with the corresponding encoding coefficients in a relatively intuitive way. Through this construction method, we are able to use the visual target and background information to construct an effective nonlinear classifier as the observation model and the nonlinear classifier will have better performance than the linear classifiers mentioned before. During the tracking process, the target and the background environment will have great changes, the training of the classifier will be changed accordingly. At this point, if the tracking failure occurs, the trained classifier will lead the target to the wrong direction, which is called the target drift (drifting). In this paper, we train an online-SVM classifier with the selected reliable training samples as an online detector, when tracking failure occurs, the detector will be activated to relocate the target in the whole image. Experiments on a benchmark demonstrate the feasibility of our method. We also give qualitative and quantitative analysis to prove the good performance of our method.